We build simulation engines for human behavior.
The name is about what simulations give you that reality doesn't, which is that history only ran once and every dataset anyone has is just a record of the version that happened, so there is no way to look up what would have gone on if the price had been ten percent lower or a different store had stocked it first. If you run the same population a few thousand times with small changes, most of those runs end up agreeing with each other, and then at certain points they stop agreeing. Those points are where the outcome is still open, and running inside them is how we extract the behavioral data that does not exist anywhere else.
Our thesis is that the most important problems are social. Arrow's information paradox, the prisoner's dilemma, the tragedy of the commons. Physical, technical, legal, are all downstream and functions of social problems. War is a stag hunt; hunger is a distribution and coordination failure; energy is veto proliferation; disease is adoption lag and free riding; science is Arrow's paradox. If you can predict human behavior you can price it, sell against it, and eventually reverse engineer to change it. In most cases the technology works or advancing it to reach a goal is a function of time and effort; the capital is available, and the legal path exists.
A few kinds, and roughly what they cost.
Information asymmetry. Two parties are each holding half of something valuable and neither one ever finds out. A cancer lab has the assay that a pharma program has been stuck on for two years, and they will never meet. What makes this hard is that neither side can go looking without giving away the thing they were trying to sell.
Adoption lag. It takes roughly seventeen years for a medical finding to reach ordinary practice. The evidence is not the constraint. The paper was published, indexed, and available to every hospital in the country the whole time.
Opposition. Nuclear plants, transmission lines, housing, datacenters, mines, etc. Developers spend nine figures on projects that get killed in hearings, and unlike a bad investment, a dead project is a straight write-off. The US interconnection queue is currently measured in terawatts and in years.
Coordination failure. Everybody would be better off if the group moved, and nobody wants to be the one who goes first. Pretty much every standard, treaty, and piece of shared infrastructure that doesn't exist, doesn't exist for this reason.
In the short term, how much money goes into decisions with incomplete information.
Imagine a beverage company about to commit a production line. They tested the flavor, it did fine, so they go national. It works in three metros and dies in the other twenty. The flavor was never really the variable. In the cities where it worked, the first few hundred people who bought it happened to be visible to everybody else, and in the cities where it died they weren't. There is currently no way to check that in advance, so they find out after the line is built.
Or take pharma. A drug launch runs on a few hundred reps calling on as many doctors as they can physically reach; physicians switch when the physicians they respect switch. If you knew which forty cardiologists the specialty was actually watching, and in what order they'd have to move, you wouldn't need the three hundred reps.
Remember Quibi? 1.75 billion dollars, gone in six months. Everybody involved was serious and every decision had research behind it. The thing none of that research could tell them was what people were going to say about it to each other in the first two weeks, which turned out to be the whole ballgame.
We think this is all one problem, and it doesn't much care what's being sold. Whatever decides if a drink ends up in a fridge is the same thing that decides if a reactor gets permitted, or if a treatment that's been proven for eleven years actually shows up in hospitals. A lot of that was invented ages ago and is just sitting there. So we want to get very good at the commercial version, partly because it's a large business and partly because as far as we can tell it's the only way anybody gets good at the other one.
Artificial populations. Each person gets a history, relationships, incentives, beliefs, and a set of people they imitate. A customer introduces something new: a product, a price, a policy, a project, a candidate. Then we run it.
Say you are launching a drug. You want to know which twelve prescribers make the difference, whether the second wave arrives at month four or month nineteen, and what happens if the wrong specialty adopts first. We tell you who moves first, who makes the decision credible, where it stalls, what becomes contagious, and what would have to change to get a different result.
Every model in commercial use averages people together and fits a curve to previous data. It averages people together and fits a curve to history. The event worth predicting only happens when people stop acting independently, so averaging removes the mechanism that produces it. Worse, these models are calibrated on the stretches where nothing happened, which makes them confidently wrong at the moment they matter.
Consumer goods and retail first, mostly because the budget is already sitting there. Every launch and price change gets tested today with surveys (where people misreport) and focus groups of twelve, and then the company commits a production line off the back of that. The shifts that actually matter now are social rather than individual — drinking is down across a generation, GLP-1s are cutting into snacking, cannabis is substituting for beer — and none of these can be surveyed, because nobody experiences their own consumption as a trend. They also cannot be tested small, since the whole thing is about what other people are doing. Market research is a $100B+ industry and most of that spend goes toward figuring out what customers did last quarter.
Pharma is the purest version. A drug's commercial outcome is decided by prescriber adoption, and physicians copy the physicians they respect. Launch strategy, opinion leader selection, and formulary sequencing are all diffusion problems, solved today with intuition and headcount.
Infrastructure developers need to know which coalition makes a project survivable before they spend the money. Utilities forecast rooftop solar, batteries, heat pumps, and EVs with curves that cannot represent contagion, then file those forecasts with a regulator. Insurers price behavior on tables built from a past that has stopped predicting the present. Funds take directional positions on what a population does next. Defense organizations model deterrence, which is a belief problem, with consultancies and slide decks.
There is a way to sort decisions that we have found useful, which is to ask two questions about any of them. First, can you test it small, and does the small test tell you anything. Second, does the outcome depend on people reacting to each other or on people deciding on their own.
If you can test it small you should, and we are worse than reality at answering it. Amazon can run a pricing experiment on four million users by Tuesday and get a cleaner answer than any simulation would give them. Same for most of e-commerce, most of app design, and anything with a fast feedback loop, i.e. the entire category the last fifteen years of data science was built around.
The interesting decisions are the ones where the small test does not generalize, which happens whenever the outcome is about what other people are doing. A pilot program in three districts tells you what happened in three districts. It does not tell you whether the thing catches, since catching is a property of the whole population and not of any sample from it. This is why pilots keep succeeding and rollouts keep failing, and why nobody in pharma or infrastructure or policy really trusts a pilot even though they all run them.
So the decisions worth simulating are expensive, made once, and dominated by people watching each other. Product launches. Price changes at scale. Plant sitings. Restructurings. Policy. Drug launches. Basically everything where the failure costs nine figures and there was never any way to check first.
There is a second business here that we did not expect when we started, which has to do with what AI labs need next.
Synthetic data is running out of road. A model generating its own training data ends up teaching itself what it already believes, and no amount of scale fixes that, since the problem is not volume. Nothing in the loop comes from outside the model. What labs need is data about the things models are worst at, and the thing models are worst at is predicting what people will do.
This is a strange gap because it does not look like one. Ask a frontier model what a population will do after a price change and you get a confident, articulate, plausible answer, and there is no way to check whether it is any good, since nobody has the ground truth. The evals do not exist. Social prediction has no benchmark the way math and code do, which means labs are flying blind on the one capability that matters most for anything an agent does out in the world.
We generate exactly that. Every engagement produces a population, an intervention, a prediction made before the fact, and a recorded outcome. Run enough of those and you have the first real dataset of human behavioral prediction with labels attached, plus the benchmark, plus millions of counterfactual worlds that could not be collected from reality even in principle.
There is an environments business inside this too. Labs are paying a lot right now for places where agents can practice and be scored, and nearly all of the good ones are technical, since correctness is checkable in code and math and not much else. Social tasks are the obvious next frontier and nobody has an environment for them, because you cannot score negotiation or persuasion or coalition building without a population to run it against. We would have the population.
We are less sure about this one and it is further out, but the shape is worth writing down.
Most consumer AI right now automates tasks that were annoying but easy. The tasks that are genuinely hard for a person are social ones, i.e. figuring out how a room will react, whether to make the ask now or in three weeks, who to route a request through, what happens to a relationship if you say the thing. People spend an enormous amount of private thought on this and get almost no help with it, mostly because there was never any way to model it.
An engine that can tell a pharma company which forty cardiologists matter can, in principle, tell a person which two people at their company actually decide things. Whether that turns into a product or stays a party trick we do not know yet. But it is the same machinery, and consumer distribution has historically been how these things get large.
Two reasons.
Forecasting was built to answer what a person wants. Survey, segment, extrapolate. All of it treats a person as an independent draw, and a person is not an independent draw. He is watching someone. The unit of analysis has been wrong for fifty years.
The theory, meanwhile, has been finished and unused for decades. It was never installed inside anything predictive, because an agent used to be three rules and a probability table, and imitation between agents that simple produces nothing. That changed in the last two years. Nobody has picked it up.
Building the simulation is not really the hard part. Showing that it predicts anything is, and that is the part most of the field skips (there are papers arguing this openly). We would rather be useful to the people working on this, so we plan to publish our forecasts before the outcomes are known, along with the calibration data and the cases where we got it wrong.
Surveys record what people say. Transactions record what they already did. History records the single path the world took.
We keep a record of populations meeting a new decision: what was introduced, what we predicted, what people actually did, and how it turned out. We also keep the worlds that did not happen, where a different group moved first, or the price or the timing or the messenger was changed.
Nobody has this and you cannot buy it, since it is not a measurement of the world so much as a measurement of a decision nobody had made yet. Foundation models are going to need it to reason about anything social, and any serious multi-agent simulation is going to need it as ground truth. We would like to be the ones who have it.
If we can build simulation engines that predict what a population does, we can tell a country how to get its power plants built. We can tell two labs on opposite sides of the world that they have each other's missing piece. We can tell you why a treatment that works has gone eleven years without being used, and what would end that.
And the engines do not have to stop at behavior. The same setup models what a population believes, what it is afraid of, what it finds respectable, and what it thinks everyone else believes (usually the part that decides things). Once those are state, the rest is arithmetic. How many people have to switch before switching is the default. Which person moves the most others. Where a belief goes from private to public, which is the same math as a market turning or a regime falling.
There is a question I keep coming back to, which is whether you can model the part of a person that decides what feels desirable or respectable or frightening, not the soul in any religious sense but just the part that settles what somebody wants before they have thought about it at all.
Magic usually means making impossible things happen through rituals or secret methods, which sounds silly, but there is a fairly literal version of this inside human behavior. Somebody believes a thing, other people notice them believing it, that attention turns into status, status changes what people do, and eventually the changed behavior becomes the evidence that the belief was correct in the first place. An idea nobody copies stays a delusion, and the same idea copied by enough people ends up being a currency or a company or a country, without anything about the idea itself having changed.
Money works this way, and so do brands, bank runs, political movements, and honestly most institutions. If we can simulate that loop then we are not only predicting what people will do, we are learning how new realities get made.
We are trying to predict human behavior, though not for a single person (one person is mostly noise) but for a population, which turns out to be fairly regular, since people copy each other and copying has structure to it.
Once you can predict a population you can also move one. Most attempts at this go through persuasion, which is expensive and mostly does not work. We would rather change the conditions people are standing in, since desire is borrowed and the conditions are what determine who they are borrowing it from.
Asimov called the study of this psychohistory and assumed one person could not be predicted while a civilization could. We think he was right, and we would like to build the working version. We want to systematize, game, and control mimesis, and to introduce new ideas to the world for adoption. We view the world as dormant, and our wish is to awaken it.
Email: ryaanaqid at gmail dot com